Latent Attribute Control for Story Generation

Neural open-domain story generation aims to generate long and fluent text as human writing. Recent work attempts to generate stories in fine-grained controls such as plot-like structure and ending valence. Although those outputs comply with the rules of grammar, they generally have logical conflicts and a lack of long-range cohesion because of explicit controlling. In this study, we propose to capture challenging story representation using latent variable modeling for the storytelling model, and we align the encoder output with story latent embeddings. Our approach and baselines are all built on the pre-trained BART language model. Experimental results demonstrated that our model largely improved compared to strong baselines on human evaluation. Human evaluators favored our generated stories, and the results were more relevant to the prompt and more coherent than the baselines.